NovaFabric — Replay and inspect what an AI agent actually did.
NovaFabric captures an agent execution as a portable Run Capsule. Teams can validate, inspect and compare runs within their own infrastructure, with optional cryptographic sealing and support for mocked replay on documented API paths.
The responsibility problem
An agent can generate a useful answer while leaving too little evidence to explain its tool calls or reproduce its behavior. NovaFabric focuses on portable execution records rather than treating a live dashboard as the only source of truth.
What you can inspect
The public project documents command capture, validation, forensic replay, structural diff and provenance. Supported mocked replay reuses recorded model and MCP results; other tools can still execute live, so the limitations matter.
Today, and what is not established
- Capture of an agent execution as a portable Run Capsule
- Validation of a capsule
- Forensic replay, structural diff and provenance
- Supported mocked replay that reuses recorded model and MCP results
- Optional cryptographic sealing (needs a configured signing key)
- Proof that an agent was correct, harmless or compliant
- A substitute for live monitoring and alerting
- Mocked replay of every tool: other tools can still execute live
Research boundary
A portable record does not itself prove that the agent was correct, harmless or compliant. Cryptographic sealing is optional and requires a configured signing key. NovaFabric is not a substitute for live monitoring and alerting.
Evidence and links
- Public sourceThe open-source repository
- Project documentationThe project’s own site
- Research preprintarXiv 2609.12582
Each link goes to the project’s own channels. This page explains why the work fits the lab; the documentation is the place to learn how to use it.
Explore the other research projects
Each project has its own authors, evidence, code availability and maturity level.
View all projects